Skip to content
FrontierPicks

Anatomy · how a dossier is built

Every dossier, dissected.

A dossier isn't a stock tip — it's a structured argument you can audit line by line. Here's a live one, taken apart: every part, what it claims, and which stage and frontier model wrote it.

The specimen

We're dissecting AMD — Advanced Micro Devices, Inc., a live dossier from the book. The specific name rotates as the book does; the structure below never changes — every one of the 929 dossiers is built from these same parts.

  1. Header — ticker, name, status

    Stage 01 Rules-based

    Identity and lifecycle state. A name only earns a dossier after it clears the tradability gate.

    AMD → AMD · Advanced Micro Devices, Inc. · Held

    How the model produces this

    The screen filters a ~400-name universe by spread, ADV, market cap and shape. Status (Held / Watchlist / Rotation / Recently exited / Dormant) tracks where the name sits in the research lifecycle — rules-based, no model judgement.

  2. Current thesis

    Stage 03 Claude Opus

    One paragraph: what the bet is, and why — in plain English.

    AMD → Advanced Micro Devices’ Helios thesis depends on an operating OpenAI deployment in the fourth quarter of 2026. September 9 analyst target changes extend favorable coverage, but delayed activation or a weekly close below $456 would…

    How the model produces this

    In the assessment stage, Claude Opus reads 30 days of news, four quarters of earnings transcripts, and filings for the name, then compresses the whole picture into a single falsifiable claim. Not a price target — a reason.

  3. Invalidation trigger

    Stage 03 Claude Opus

    The explicit kill criterion — published up front, before it fires.

    AMD → A weekly close below $456 invalidates the September recovery by undercutting the 2026-09-03 market close of $456.16; independently, AMD or OpenAI postponing Helios activation beyond Q4 2026 breaks the deployment…

    How the model produces this

    The honesty mechanism. Opus commits, in writing, to the condition that would prove the thesis wrong. If it triggers, conviction is stripped and the thesis is treated as broken. A published commitment, not a soft caveat.

  4. Bull case · Bear case

    Stage 03 Claude Opus

    Both sides, constructed identically. No hand-waving on the side it disagrees with.

    AMD → Present — sourced, dated, equal weight

    How the model produces this

    Opus writes the bear case with the same rigour as the bull case — sourced bullets, dated, weighted. A dossier you can audit shows the disconfirming evidence, not just the conclusion.

  5. Setup & price structure

    Stage 02 Rules-based + Sonnet

    Where the name is technically — the structure the thesis has to play out across.

    AMD → Present — MAs, RSI, levels, basing pattern

    How the model produces this

    The scoring stage weighs price structure × volume × news density × theme-cluster strength. The structural read (moving averages, RSI, key levels) feeds the assessment so the thesis is grounded in actual price behaviour, not just the narrative.

  6. Catalyst calendar

    Stage 03 Claude Opus

    The dated events that could move the thesis, next 30 days.

    AMD → Next dated catalyst · 2026-09-11

    How the model produces this

    Earnings prints, product dates, regulatory decisions — extracted and dated so a reader knows exactly when the binary resolves. Dossiers age fast; the calendar is why recency-of-write is on every page.

  7. What would change our mind

    Stage 03 Claude Opus

    The conditions that would raise or lower conviction — stated, not implied.

    AMD → Present — explicit higher / lower conviction conditions

    How the model produces this

    Separate from the binary kill-switch: the graded evidence that would move confidence up or down. Makes the model's reasoning legible — you can see what it's waiting for.

  8. Archetype

    Stage 03 Claude Opus + sizing rules

    A behaviour profile that hard-caps how big a name can get — not a label.

    AMD → Cyclical recovery

    How the model produces this

    One of eight archetypes (Compounder → Macro hedge). The sizing rules read it directly: a Retail-squeeze name gets a hard size cap a Compounder never would. See the full taxonomy.

    Archetype taxonomy →
  9. Conviction

    Stage 03 Claude Opus

    Calibrated confidence that the setup plays out — not a return forecast.

    AMD → MEDIUM

    How the model produces this

    Four tiers (Supreme → Low). Each calibrates sizing and stop discipline, and each carries its own invalidation trigger that strips it if breached. Confidence with a published off-ramp.

    conviction highmediumlownone

  10. Themes & correlation

    Stage 02 Rules-based + Sonnet

    The narrative baskets the name belongs to, and how it moves with them.

    AMD → Mega-cap AI platforms

    How the model produces this

    Themes are treated as primary. The scoring stage reads basket behaviour — a name is sized against its cohort and the regime, not in isolation. Correlation notes capture how it actually trades with the group.

    Browse this theme →

Three mechanisms that keep it honest

Invalidation

Every thesis ships with the condition that would prove it wrong — published before it fires.

Dating

Every read carries its write date. A confident setup read can age out in hours; the page tells you when it was made.

Symmetry

The bear case is built with the same rigour as the bull case. Disconfirming evidence is shown, not buried.